Nonlinear Signal Analysis For Rolling Bearing Condition Monitoring And Fault Diagnosis

Considering the nonlinearity or nonlinear components contained in vibration signals from mechanical equipment, this paper investigates nonlinear signal analysis to extract the nonlinear feature from vibration signals, and addresses on its applications to condition monitoring and faults diagnosis of rolling bearing. Four nonlinear signal analysis methods including complexity measurement—permutation entropy (PE), kernel principal component analysis (KPCA), manifold learning and support vector machine (SVM) are used to extract the nonlinear features from vibration signals of rolling bearing. And applications of the four methods in rolling bearing condition monitoring and faults diagnosis are investigated in detail.Permutation entropy, as one of the complexity measurements, is an important nonlinear feature which can represent disorder status of signals. As a feature representing bearing status, PE was investigated to detect the dynamic changes of vibration signals of rolling bearing. Through being tested by the classic nonlinear time series—Logistic Map, PE was validated to be sensitive to dynamic changes of nonlinear signal and enable detecting the changes effectively. Through analyzing effect of data length, delay time, embedded dimension and time cost, we selected the optimal parameters which fit to vibration signal of rolling bearing and calculated the PEs of rolling bearing on different statuses. By identifying changes in the vibration signals measured on rolling bearing, which are typical precursors of defect occurrence, permutation entropy can serve as a diagnostic tool. Experiments on a custom-designed gearbox system have confirmed its effectiveness for bearing health monitoring applications.Taking advantage of performance on nonlinear data mining and dimension reduction of kernel principal component analysis (KPCA), this paper address on extracting the novel and more sensitive nonlinear features from the original time-domain, frequency-domain and time-frequency domain features to represent and classify rolling bearing conditions based on KPCA. By conducting the KPCA on four conditions of the bearing, the extracted nonlinear features have better performance in clustering and status representation than other traditional methods. Therefore, the nonlinear features can be used to identify health status of rolling bearing. Furthermore, Hotelling's T2 and Q-statitic are developed to monitor defect status of rolling bearing. And Experiments on a bearing system have confirmed its effectiveness for monitoring defect status of bearing.A novel method of defect detection and diagnosis based on manifold learning was proposed in this paper and used to extract nonlinear feature in low dimensional manifold from original statistic features in high dimension. This paper investigates the method of representing bearing health statuses using the nonlinear feature extracted by manifold learning. Three manifold learning algorithms including LLE, ISOMAP and LTSA, are investigated to extract nonlinear features from simulated signals of bearing and Swiss Roll and Swiss Hole data. The results show that LTSA algorithm can effectively extract nonlinear features from high dimensional data and the nonlinear features have good performance in clustering and representing status of bearing. By analyzing the nonlinear features extracted from vibration signals of rolling bearing on four bearing conditions using LTSA, the extracted nonlinear features have better clustering effect and smaller class-to-class distance than the features extracted by PCA, and it can be used to represent and recognize rolling bearing statuses.Pattern Recognition is essential to faults diagnosis of rolling bearing. In view of lacking samples of fault data and nonlinearity of bearing vibration, this paper addressed on fault diagnosis of rolling bearing using SVM and investigated to construct the SVM multi-class classifier based on"one against all"algorithm. By selecting optimal parameters using cross validation method, the multi-class SVM classifier was constructed based on gauss radial kernel function. And study on simulating signals has confirmed that the SVM classifier has good performance on classification. Combination of KPCA and LTSA with SVM was investigated to recognize the conditions of rolling bearing. Results show that the multi-class SVM classifier is an effective and efficient tool for identifying rarely defect conditions of rolling bearing for its high correct recognizing rate, low training time and other merits.Moreover, the research methods described above were all validated by experimental analysis. Vibration signals of rolling bearing were acquired using data acquisition system constructed by virtual instrument technology. Based on rotating machine test rig and bearing-gearbox, experiments with seeding faults on rolling bearing were carried out to verify the feasibility and the effectiveness of the methods above. The study of this paper indicates that nonlinear signal analysis can effectively extract nonlinear features and identify defect conditions of bearing, which is very significant to condition monitoring and faults diagnosis of rolling bearing.